RoentGen-v2: Demographically Controlled CXR Synthesis

• The paper introduces RoentGen-v2, a demographically controllable text-to-image latent diffusion model that synthesizes PA-view chest radiographs using explicit radiographic and demographic prompts.
• It employs a two-stage training pipeline with synthetic pretraining on demographically balanced data to improve classifier performance, robustness to distribution shifts, and subgroup fairness.
• RoentGen-v2 leverages Stable Diffusion v2.1 architecture with fine-tuning on image-report-demographics triplets and strict quality control, enabling construction of large-scale balanced synthetic datasets.

Searching arXiv for the cited papers to ground the article and citations. RoentGen-v2 is a demographically controllable text-to-image latent diffusion model for chest radiography introduced in "Improving Performance, Robustness, and Fairness of Radiographic AI Models with Finely-Controllable Synthetic Data" (Moroianu et al., 22 Aug 2025). It enables fine-grained conditioning on both radiographic findings and patient demographic attributes—sex, age, and race/ethnicity—and uses that controllability to construct a large, demographically balanced synthetic chest radiograph dataset. The system is positioned not merely as a generative model, but as a data-centric mechanism for improving downstream disease classification under limited and imbalanced labeled data. In the reported evaluation, its principal effect is achieved through supervised pretraining on synthetic data followed by fine-tuning on real data, yielding simultaneous gains in in-distribution performance, out-of-distribution generalization across institutions, and subgroup fairness.

1. Definition, scope, and central contribution

RoentGen-v2 upgrades the original RoentGen model by fine-tuning Stable Diffusion v2.1 on paired CXR–report–demographics triplets, thereby adding demographic control alongside finding-level control (Moroianu et al., 22 Aug 2025). The resulting model can synthesize clinically plausible PA-view chest radiographs while explicitly setting sex, age, and race/ethnicity in the prompt. The stated demographic variables are sex in $\{ \text{male}, \text{female} \}$, race/ethnicity in $$\{ \text{Asian}, \text{Black}, \text{Hispanic}, \text{White} \}$$, and numeric age.

The central goal is to improve model performance, robustness to distribution shift across institutions, and fairness across demographic subgroups, while operating under real-world constraints of limited and imbalanced labeled data. The key innovation is explicit demographic conditioning in generation, which permits targeted balancing when the real data distribution is skewed, including long-tailed race categories, and supports systematic fairness analysis and mitigation through synthetic pretraining rather than simplistic mixing of data sources.

A common source of confusion is the role played by the synthetic data. RoentGen-v2 is not presented as an end in itself; it is used to build a balanced synthetic dataset for training downstream disease classifiers. The reported contribution therefore spans both generation and a downstream training protocol. In this framing, the model’s importance lies in its capacity to produce controlled cohorts for classification pretraining, rather than only in unconditional or weakly controlled image synthesis.

2. Generative architecture and conditioning formulation

RoentGen-v2 uses Stable Diffusion v2.1 as its base architecture: a latent diffusion model with a U-Net denoiser in latent space, a VAE image encoder/decoder, and CLIP ViT-L/14 text embeddings with a 77-token limit (Moroianu et al., 22 Aug 2025). Conditioning text is composed from radiology report impressions and demographic metadata, and the CLIP text encoder and U-Net are jointly fine-tuned. Cross-attention throughout the U-Net integrates text conditioning.

The prompt template is structured as:

"<AGE> year old <RACE> <SEX>. <IMPRESSION>"

If the impression exceeds the 77-token CLIP limit, the impression is summarized using GPT-4 to $\le 200$ characters; 6,072 impressions were summarized. Findings are controlled via the impression section of the report. Common labels include Atelectasis, Cardiomegaly, Consolidation, Edema, Effusion, Pneumonia, Pneumothorax, and "No Finding." The model supports compositional control, so multiple findings plus demographic attributes can be combined within a single prompt.

The training corpus is a curated subset of MIMIC-CXR v2.0.0 containing PA-view images, impression sections, and demographics. After exclusion, the dataset contains 75,639 studies total, with 66,760 for training, 1,295 for validation, and 7,584 for test; approximately 68k image–report pairs were used for RoentGen-v2 training, compared with approximately 35k for RoentGen-v1. Optimization uses AdamW, a learning rate of $5 \times 10^{-5}$, constant learning rate with 500-step warm-up, batch size 192, and training up to 60k steps (172 epochs). Checkpoint selection favored 10k steps (approximately 28 epochs), requiring approximately 14 A100 (40GB) GPU-hours; full 60k-step training took approximately 80 GPU-hours.

The diffusion formulation follows the standard noise-prediction objective in latent space:

$$x_t = \sqrt{\bar{\alpha}_t} x_0 + \sqrt{1 - \bar{\alpha}_t}\,\epsilon,\qquad \epsilon \sim \mathcal{N}(0, I)$$

with cumulative noise schedule $\bar{\alpha}_t = \prod_{s=1}^t \alpha_s$, using the default scheduler of Stable Diffusion v2.1. The training loss is

$$L = \mathbb{E}_{t,\epsilon,c} \left[ \lVert \epsilon - \epsilon_\theta(x_t, t, c) \rVert^2 \right],$$

where $c$ denotes the conditioning from the CLIP text embedding of the demographics-plus-findings prompt. During sampling, classifier-free guidance is used:

$$\hat{\epsilon} = (1 + w)\epsilon_{\text{cond}} - w\epsilon_{\text{uncond}},$$

with guidance scale $w = 4.0$. Unconditional inputs are obtained by dropping conditioning with a fixed probability during training.

3. Synthetic dataset construction, balancing, and quality control

Using RoentGen-v2, the authors constructed a synthetic dataset of 565,154 PA-view chest radiographs at $$\{ \text{Asian}, \text{Black}, \text{Hispanic}, \text{White} \}$$0 resolution from 623,712 prompts (Moroianu et al., 22 Aug 2025). The prompts were generated by enumerating all sex $$\{ \text{Asian}, \text{Black}, \text{Hispanic}, \text{White} \}$$1 race/ethnicity combinations for each real impression and sampling age uniformly within $$\{ \text{Asian}, \text{Black}, \text{Hispanic}, \text{White} \}$$2 years of the original age. Inference used 75 denoising steps, guidance scale 4.0, and the default Stable Diffusion v2.1 scheduler.

A dedicated QC module was introduced to enforce instruction-following for demographics via independent pretrained XRV demographic classifiers. A sample passes QC if sex and race predictions match the prompt exactly and age prediction is within $$\{ \text{Asian}, \text{Black}, \text{Hispanic}, \text{White} \}$$3 years of the prompted age, with the $$\{ \text{Asian}, \text{Black}, \text{Hispanic}, \text{White} \}$$4 threshold chosen to match the XRV age RMSE standard deviation on real data. Failed prompts are regenerated up to three times, and persistent failures are discarded. Of all prompts, 58,558 (9.4%) failed QC; 9.2% of failures were due to race, 0.2% due to age, and none due to sex. The reported interpretation is that this pattern reflects the weak direct imaging correlates of race and imbalance in the generative training distribution.

The demographic balancing achieved by the synthetic set is summarized below.

Group	Real count	Synthetic count
White	45,102	155,728
Black	13,685	155,653
Hispanic	5,521	155,870
Asian	2,452	97,903

This balancing is central to the downstream fairness analysis. Real medical imaging datasets are commonly imbalanced across demographic attributes, and the paper explicitly links such imbalance to reduced generalization and to fairness gaps in underrepresented groups. By generating balanced cohorts across sex, age, and race/ethnicity, RoentGen-v2 provides a mechanism for equitable representation during training and for explicit subgroup evaluation.

The synthetic dataset has 83% samples with one or more positive findings and 17% "No Finding." Generation plus QC required approximately 360 A100 (40GB) GPU-hours. Fidelity and diversity were assessed at the 10k-step checkpoint. Prompt–label alignment on synthetic images yielded average disease AUROC 0.81, compared with a real baseline XRV disease AUROC of 0.88; sex accuracy was 100% versus 97% on real data; race accuracy was 98% versus 95%; age RMSE was 8.9 years versus 7.1 years. FID was 76.8, compared with 96.1 for RoentGen-v1. Diversity for repeated sampling from the same prompt was characterized by MS-SSIM 0.37 and BioViL embedding cosine 0.66, which the paper interprets as indicating non-collapsed variability. No radiologist review was reported; plausibility was inferred via automated metrics and exemplar visualizations.

4. Downstream classification pipeline and training strategies

The downstream task is multi-label classification of the 14 CheXpert classes, consisting of 13 findings plus "No Finding," with evaluation focused on eight labels common across datasets: Atelectasis, Cardiomegaly, Consolidation, Edema, Effusion, Pneumonia, Pneumothorax, and No Finding (Moroianu et al., 22 Aug 2025). The classifier architecture is DenseNet-121, selected based on prior chest radiograph classification performance. Preprocessing consists of center-crop, resize to $$\{ \text{Asian}, \text{Black}, \text{Hispanic}, \text{White} \}$$5, and normalization using mean and standard deviation from the MIMIC training set.

Optimization uses AdamW with cosine learning-rate scheduling, initial learning rate 0.0001 for ImageNet-initialized models, weight decay 0.05, and a maximum of 100 epochs with early stopping if validation loss is stagnant for more than 20 epochs. Per-label thresholds are chosen to maximize F1 on the validation set, and macro-averaged AUROC and AUPRC are reported.

Four training strategies were compared:

1. Real-only: ImageNet initialization; training on 66k real images.
2. Synthetic-only: ImageNet initialization; training on synthetic data, up to 565k images.
3. Synthetic+Real mix: ImageNet initialization; training on 66k real plus varying synthetic sizes, up to 565k.
4. Synthetic pretraining: DenseNet-121 trained from scratch on 565k synthetic images with learning rate $$\{ \text{Asian}, \text{Black}, \text{Hispanic}, \text{White} \}$$6, then fine-tuned on real data with learning rate $$\{ \text{Asian}, \text{Black}, \text{Hispanic}, \text{White} \}$$7.

The two-stage synthetic pretraining pipeline is operationally simple. Synthetic data are generated with demographic balancing and QC, then used for supervised pretraining with multi-label BCE loss on 14 classes, after which the classifier is fine-tuned on real MIMIC-CXR data. Evaluation is performed on MIMIC for in-distribution testing and on CheXpert, NIH ChestX-ray, PadChest, and VinDr-CXR for out-of-distribution testing, totaling more than 137,000 test chest radiographs across five institutions.

This training design is the central methodological claim of the work. The paper explicitly contrasts synthetic pretraining with naïve data mixing, arguing that balanced synthetic cohorts are most effective when used to shape feature learning before exposure to real data distributions.

5. Performance, out-of-distribution generalization, and fairness

On in-distribution MIMIC test data, naïve mixing did not improve AUROC relative to the real-only baseline: synthetic+real training with 66k real plus 565k synthetic achieved AUROC 0.770 (CI 0.762–0.778), versus 0.772 (CI 0.765–0.781) for real-only, with $$\{ \text{Asian}, \text{Black}, \text{Hispanic}, \text{White} \}$$8 (Moroianu et al., 22 Aug 2025). Out-of-distribution performance with the mixed strategy was better than real-only on CheXpert, NIH, and PadChest, but not on VinDr; gains plateaued at approximately 300k synthetic samples, or about $$\{ \text{Asian}, \text{Black}, \text{Hispanic}, \text{White} \}$$9 the real training set.

Synthetic-only training showed a mixed pattern: it underperformed real-only on MIMIC, with AUROC 0.756 versus 0.772 ($\le 200$0), but outperformed real-only on CheXpert, NIH, and PadChest, while showing no significant difference on VinDr. The most consistent gains were produced by synthetic pretraining followed by real fine-tuning.

Dataset	Real-only AUROC	Synthetic pretraining + fine-tuning AUROC
MIMIC	0.772	0.798
CheXpert	0.700	0.732
NIH	0.661	0.726
PadChest	0.721	0.781
VinDr	0.641	0.683

These correspond to AUROC increases of +3.3% on MIMIC, +4.5% on CheXpert, +9.8% on NIH, +8.3% on PadChest, and +6.5% on VinDr, all with $\le 200$1. AUPRC improved as well, including 0.360 versus 0.316 on MIMIC, 0.188 versus 0.160 on NIH, 0.195 versus 0.156 on PadChest, and 0.258 versus 0.225 on VinDr, all with $\le 200$2. The headline summary is an average 6.5% AUROC increase across datasets from synthetic pretraining, compared with 2.7% from naïve mixing.

The paper also reports strong data-efficiency effects. With synthetic pretraining, fine-tuning on fewer than 10k real images matches or exceeds the real-only model trained on 66k real images. With at least 30k real images, synthetic-pretrained models surpass the synthetic+real mixed model. This suggests that the main effect of the synthetic data is representational initialization rather than mere sample-count augmentation.

Fairness is quantified using two metrics. For subgroup parity, the AUROC fairness gap is

$\le 200$3

For underdiagnosis, the analysis focuses on the "No Finding" label. For subgroup $\le 200$4,

$\le 200$5

and the underdiagnosis fairness gap is

$\le 200$6

Lower $\le 200$7 indicates more equitable missed-diagnosis rates across subgroups. Subgroups were evaluated intersectionally as sex $\le 200$8 race/ethnicity for MIMIC and CheXpert, and sex $\le 200$9 age bins for MIMIC, CheXpert, NIH, and PadChest.

Synthetic pretraining reduced fairness gaps while increasing AUROC. Average out-of-distribution reduction in AUROC fairness gap was 16.0% with +7.3% AUROC. Average underdiagnosis gap reduction across datasets was 19.3% with +6.5% AUROC. Reported examples include sex $5 \times 10^{-5}$0 race/ethnicity underdiagnosis gap reductions from 0.260 to 0.210 on MIMIC and from 0.123 to 0.099 on CheXpert, and sex $5 \times 10^{-5}$1 age reductions from 0.614 to 0.519 on MIMIC, 0.434 to 0.346 on CheXpert, 0.349 to 0.269 on NIH, and 0.603 to 0.490 on PadChest. Calibration and uncertainty analyses were not reported.

6. Relation to prior work, limitations, and practical interpretation

Relative to RoentGen-v1, RoentGen-v2 trains on approximately twice as many image–report pairs, upgrades from Stable Diffusion v1.4 to v2.1, adds demographic conditioning, and improves FID from 96.1 to 76.8 while maintaining comparable disease alignment, 0.81 versus 0.82 AUROC (Moroianu et al., 22 Aug 2025). The paper also situates diffusion models against GAN and VQ-VAE approaches, stating that diffusion models provide higher fidelity, better prompt adherence, and more stable training across medical imaging modalities. In the specific setting of synthetic augmentation, the reported contribution is not that synthetic data universally help when mixed into real training, but that synthetic pretraining is superior to naïve mixing for both performance and fairness.

RoentGen-v2 should also be distinguished from contemporaneous radiology-language systems. "R2GenKG: Hierarchical Multi-modal Knowledge Graph for LLM-based Radiology Report Generation" addresses hallucination and weak disease diagnostic capability in report generation through a multimodal knowledge graph, R-GCN encoding, disease-aware vision token retrieval, and Llama2-7B decoding (Wang et al., 5 Aug 2025). "Act Like a Radiologist: Radiology Report Generation across Anatomical Regions" addresses multi-region report generation through region-aware knowledge aggregation and implicit prior guidance (Chen et al., 2023). Those systems concern report generation, whereas RoentGen-v2 concerns controllable chest radiograph synthesis and the use of synthetic images to train disease classifiers. A plausible implication is that these lines of work are adjacent rather than interchangeable: one governs image generation for downstream classification, the others govern text generation from images.

The principal limitations are explicit. RoentGen-v2 is trained from a single source, MIMIC-CXR, and only on PA-view chest radiographs. The Asian subgroup had higher QC failure due to fewer real training examples and weak imaging correlates for race. Synthetic prevalence mirrored real training distributions, so long-tailed diseases were not oversampled. Labels are report-derived and may contain noise, which can affect fairness metrics. Although the output is synthetic, the paper notes privacy risks from possible memorization by diffusion models and calls for further privacy audits and memorization safeguards. It also states that demographic conditioning can be misused and that governance should restrict use to fairness-oriented research and monitored clinical development, with appropriate IRB and data-sharing policies.

The usage guidance is correspondingly specific. Synthetic pretraining is recommended for disease classifiers in data-scarce or imbalanced settings, especially when out-of-distribution deployment and subgroup fairness are priorities. The suggested procedure is to generate demographically balanced synthetic cohorts with strict QC, pretrain on a large synthetic set—300k images is identified as sufficient for out-of-distribution gains when compute is limited—and then fine-tune on available real data. Evaluation should jointly monitor overall performance and subgroup metrics using $5 \times 10^{-5}$2 and $5 \times 10^{-5}$3, rather than treating fairness and performance as inherently opposed objectives. This suggests a data-centric interpretation of RoentGen-v2: its main contribution is to show that finely controllable synthetic chest radiographs can serve as an effective pretraining substrate for more robust and more equitable medical imaging models.